perf bench sched cpu-matrix benchmark

Hi,

During the discussion of the nohz load balancer fix
(https://lkml.org/lkml/2011/9/27/197), Ingo Molnar suggested to replace the simple while1 loop, used for publishing the test results with cpu-matrix test. This patch adds cpu-matrix benchmark under perf bench.

TODO:
- Provide option to run main thread at real-time priority.
- Provide option for changing scheduling priority/policy of worker threads.

perf bench: Add sched cpu-matrix benchmark

perf bench sched cpu-matrix benchmark is a matrix multiplication workload, which can be replaced with the traditional while1 cpu hog.

Example of usage:

% perf bench sched cpu-matrix
# Running sched/cpu-matrix benchmark...

Multiplication of [20] x [20] matrix, using [1] threads
Total time: 0.000170 [sec]

% perf bench sched cpu-matrix -s1k -t3
# Running sched/cpu-matrix benchmark...
elapsed time = 0.100066 sec, progress = 27270868
elapsed time = 0.100079 sec, progress = 27299841
elapsed time = 0.100071 sec, progress = 27253513
elapsed time = 0.100073 sec, progress = 27272693
elapsed time = 0.100090 sec, progress = 10208415

Multiplication of [1024] x [1024] matrix, using [3] threads
Total time: 0.500495 [sec]

% perf bench --format=simple sched cpu-matrix -s1k -t3
Total time: 0.454764 [sec]

--

 tools/perf/Makefile           |    1 +
 tools/perf/bench/bench.h      |    2 +
 tools/perf/bench/cpu-matrix.c |  410 +++++++++++++++++++++++++++++++++++++++++
 tools/perf/builtin-bench.c    |    3 +
 4 files changed, 416 insertions(+), 0 deletions(-)

index b98e307..02bd562 100644
@@ -363,6 +363,7 @@ ifeq ($(RAW_ARCH),x86_64)
 BUILTIN_OBJS += $(OUTPUT)bench/mem-memcpy-x86-64-asm.o
 endif
 BUILTIN_OBJS += $(OUTPUT)bench/mem-memcpy.o
+BUILTIN_OBJS += $(OUTPUT)bench/cpu-matrix.o
 
 BUILTIN_OBJS += $(OUTPUT)builtin-diff.o
 BUILTIN_OBJS += $(OUTPUT)builtin-evlist.o

index f7781c6..174465a 100644
@@ -4,6 +4,8 @@
 extern int bench_sched_messaging(int argc, const char **argv, const char *prefix);
 extern int bench_sched_pipe(int argc, const char **argv, const char *prefix);
 extern int bench_mem_memcpy(int argc, const char **argv, const char *prefix __used);
+extern int bench_cpu_matrix(int argc, const char **argv,
+				const char *prefix __used);
 
 #define BENCH_FORMAT_DEFAULT_STR	"default"
 #define BENCH_FORMAT_DEFAULT		0

index fcb9626..df84428 100644
@@ -42,6 +42,9 @@ static struct bench_suite sched_suites[] = {
 	{ "pipe",
 	  "Flood of communication over pipe() between two processes",
 	  bench_sched_pipe      },
+	{ "cpu-matrix",
+	  "Benchmark to run cpu matrix multiplication",
+	  bench_cpu_matrix      },
 	suite_all,
 	{ NULL,
 	  NULL,

Fix typo:
s/multipilcations/multiplications

We should perhaps add 'sleep_msec' argument to this function.

- vatsa

It would be nice to also print the *size* of the matrix, in MBs or so. 1024x1024 of int, long, float ... or double?

Looks pretty useful otherwise. Arnaldo, any objections?

Thanks,

Ingo

Hi Ingo,
Thanks for the review.

Changes from v1:
- Addressed the review comments from Vatsa and Ingo.
- Added support for creating master/worker thread(s) with individual policy and priority.
- Changed the format of printing the progress multiplications of thread(s).

perf bench: Add sched cpu-matrix benchmark

perf bench sched cpu-matrix benchmark is a matrix multiplication workload, which can be replaced with the traditional while1 cpu hog.

Example of usage:

% perf bench sched cpu-matrix
# Running sched/cpu-matrix benchmark...

Multiplication of [20] x [20] matrix, using [1] threads
Total time: 0.000170 [sec]

% perf bench sched cpu-matrix -s1024 -t10 -p1000
# Running sched/cpu-matrix benchmark...
57701987 multiplications over 1.000069 [sec]
60361191 multiplications over 1.000119 [sec]
...
20753139 multiplications over 1.000151 [sec]

Multiplication of [1K] x [1K] matrix, of unsigned int using [10] threads
Total time: 19.002301 [sec]

% perf bench --format=simple sched cpu-matrix -s1k -t10
Total time: 18.601030 [sec]

% perf bench sched cpu-matrix -s1k -t10 -wr99 -p1000 # Running sched/cpu-matrix benchmark... 56894683 multiplications over 1.000071 [sec] 53085285 multiplications over 1.000115 [sec]
...
37727908 multiplications over 1.000127 [sec]

Multiplication of [1K] x [1K] matrix, of unsigned int using [10] threads
Total time: 19.002302 [sec]

---
 tools/perf/Makefile           |    1 +
 tools/perf/bench/bench.h      |    2 +
 tools/perf/bench/cpu-matrix.c |  573 +++++++++++++++++++++++++++++++++++++++++
 tools/perf/builtin-bench.c    |    3 +
 4 files changed, 579 insertions(+), 0 deletions(-)

index b98e307..02bd562 100644
@@ -363,6 +363,7 @@ ifeq ($(RAW_ARCH),x86_64)
 BUILTIN_OBJS += $(OUTPUT)bench/mem-memcpy-x86-64-asm.o
 endif
 BUILTIN_OBJS += $(OUTPUT)bench/mem-memcpy.o
+BUILTIN_OBJS += $(OUTPUT)bench/cpu-matrix.o
 
 BUILTIN_OBJS += $(OUTPUT)builtin-diff.o
 BUILTIN_OBJS += $(OUTPUT)builtin-evlist.o

index f7781c6..174465a 100644
@@ -4,6 +4,8 @@
 extern int bench_sched_messaging(int argc, const char **argv, const char *prefix);
 extern int bench_sched_pipe(int argc, const char **argv, const char *prefix);
 extern int bench_mem_memcpy(int argc, const char **argv, const char *prefix __used);
+extern int bench_cpu_matrix(int argc, const char **argv,
+				const char *prefix __used);
 
 #define BENCH_FORMAT_DEFAULT_STR	"default"
 #define BENCH_FORMAT_DEFAULT		0

index fcb9626..df84428 100644
@@ -42,6 +42,9 @@ static struct bench_suite sched_suites[] = {
 	{ "pipe",
 	  "Flood of communication over pipe() between two processes",
 	  bench_sched_pipe      },
+	{ "cpu-matrix",
+	  "Benchmark to run cpu matrix multiplication",
+	  bench_cpu_matrix      },
 	suite_all,
 	{ NULL,
 	  NULL,